The Shift From Answers to Actions

Hyperscalers are projected to spend nearly $800 billion on AI infrastructure in 2026, and markets are already asking where the returns are. Railways, electricity and fiber were all overbuilt before demand caught up, and AI looks set to follow. What comes next is specialist AI agents that deliver measurable outcomes, and the trust they need to operate at scale.

AI Infrastructure and Timing 

Is an AI infrastructure correction likely? 

Yes. Hyperscalers are projected to spend nearly $800 billion on capital expenditure in 2026, with higher forecasts for 2027, while markets are already asking where the returns are. Major general-purpose technologies that require heavy infrastructure have historically been overbuilt ahead of demand. Technology can still be real even when the financing is wrong; railways and the internet both produced lasting prosperity after their respective crashes. 

What historical pattern is AI following? 

Infrastructure first, native applications later. Major technological revolutions typically move through three stages: 

1. Infrastructure and technical standards (power grids, fiber, GPU clusters and foundation models). 

2. Replacement applications that improve known activities (electric light replacing gas lamps; chatbots replacing search and basic drafting). 

3. Native applications that only make sense once the new infrastructure exists (Uber, Airbnb, TikTok). 

Stage 3 native AI applications are not yet clearly visible. Paul David’s work on electrification showed that productivity gains took decades because factories first attached electric motors to steam-era layouts; real gains arrived only after redesigning around electricity itself. 

Why is the timing tighter for AI than for the fiber boom? 

GPUs have a short economic life. Hyperscalers typically depreciate servers over roughly five to six years, and many analysts argue the effective life of frontier accelerators is closer to three years as each generation resets performance per watt. Fiber could sit dark for years; AI infrastructure cannot. At the same time, the cost of intelligence at a given capability level has collapsed, making continuous multi-step agent workloads newly viable but also pressuring the value of already-installed chips. 

From Answers to Actions 

How does the shift from answers to actions change AI economics? 

Chatbots leave most of the economic work to the human; an agent that books, buys, files, reconciles, or coordinates multi-step processes competes with labor hours and process friction—a market measured in trillions rather than seats. The prompt gives way to delegation. This does not prevent a financing correction, but it can determine how quickly infrastructure finds productive traffic afterward. 

Why do agents improve utilization and ROI more than better chatbots? 

Four mechanisms matter: 

• Measurable ROI: an agent that runs a multi-step procurement process or reconciles records has a clear payback period that finance teams can budget. 

• Budget shift: agents move spend from easily cut innovation budgets into core operating budgets. 

• Higher utilization: a chat answer is seconds of inference; an always-on agent that monitors, calls tools, retries, and verifies can consume orders of magnitude more compute.

• Compounding: one agent action can trigger others (booking a shipment triggers payment, inventory update, and a new purchase order), creating network effects similar to those that emerged on the early web. 

Moats in the Agent Economy 

Is general-purpose action a durable moat? 

No. Booking, emailing, scheduling, and form-filling are becoming common capabilities across major platforms, just as answering questions became common. These generalist agents tend to be priced similarly and will compete primarily on distribution. Common-knowledge answers and actions trend toward the cost of production, which is falling rapidly. Durable value shifts toward specialized outcomes. 

Where do the real moats live? 

The strongest durable advantages are privileged access plus accountability / verified track record. 

• Proprietary data is an advantage but rarely a true moat (customers often own it; much can be licensed or approximated). 

• Encoded expertise (prompts, rules, workflows) is a head start that can be reverse-engineered. 

• Privileged access (licenses, regulatory approvals, write access to systems of record, payment rails, certified integrations) is much harder to copy. 

• Accountability and a verified history of outcomes (backed by certification, insurance, or liability) is the strongest and slowest to build. 

Specialist agents can be priced against the cost of the expert they augment rather than against token cost, supporting outcome-based pricing. The likely market shape is generalist agents as the front door that hand off to specialists. 

Can ordinary people’s expertise become valuable? 

Yes. Almost everyone has local, tacit knowledge that frontier models trained on public internet data lack. AI lowers the cost of packaging that knowledge into an agent; the person with the expertise no longer needs a full software team. An agent can reach a global audience without sleep or travel. The remaining constraints are discovery and trust. Kevin Kelly’s “1,000 true fans” logic extends from creative work to expertise, but lasting value still requires a creator who stands behind outcomes, a verified record, and discovery/payment infrastructure. 

Trust Layer and What Comes Next 

What is the missing trust layer for autonomous agents? 

Autonomous agents that allocate budgets, update records, or buy real assets need infrastructure built on four primitives: 

• Verifiable identity (distinguishing authorized agents and the humans or companies behind them).

• Programmable escrow and settlement (holding funds until a task is verifiably done). 

• Physical–digital verification (proving real-world actions such as deliveries or sustainability outcomes). 

• Open agent marketplaces (allowing agents from different builders to discover, hire, and pay one another). 

E-commerce scaled when people trusted websites with card numbers. Without an equivalent trust layer, autonomous execution stays constrained by closed platforms and bilateral trust. Blockchain’s relevant role is making identity, authority, settlement, and outcomes independently verifiable when they cross organizational boundaries. 

What should builders focus on? 

Build for utility. The parameter race is giving way to a utility race. Stage 1 (infrastructure) is largely built at enormous cost, Stage 2 (agent action) is arriving, and Stage 3 (native applications) remains to be invented—on a shorter clock and with a thinner fiscal safety net than 2000. The builders who survive a correction will inherit near-free intelligence and the rails to build on. The future belongs to agents that do not just speak but act—and can prove it—anchored in measurable economic reality through accountable expertise and verified outcomes. 

Sources: Sunny Lu (@sunshinelu24)